Hi Sileesh

Yes, using kernel density estimation is advisable when you analyze length frequency distributions (or any other type of frequency distributions) because it allows you to quickly explore and to characterise the distribution of your dataset. Two issues that should considered with this approach: a) type of kernel to be used (there is a bunch of them, i.e. Gaussian, tricube, triweight), b) Bandwith selection. I will just focus on the bandwith issue. See an example in the attached figure. I used R to create a toy dataset for a random variable with a tri-modal distribution (n = 3000 observations), that is simply the combination of three random variables with normal distribution with mean 0, 3 and 6, respectively, and SD=1 in all cases. 

When your bandwith is too high (bw = 2 or 1), the resulting frequency distribution showed by the kernels indicate a rather unimodal normal distribution. The tri-modal distribution becomes apparent only when you decrease the bandwith. Of course decreasing the bandwith too much increases the noise. I know there are some approaches to estimate the optimal bandwith, but honestly I am not familiar with them. Said that, if you are only interested in exploring length distribution data, kernel density estimation may be very useful technique in help you identifying the presence of multiple size cohorts (or the lack of them). 

Good luck with your research